A modular AI framework built from Agentic Units — self-contained microservices that discover, collaborate, and adapt.
Launch Intent StudioThree core services orchestrate a network of specialized AI agents
Translates natural language intent into a directed acyclic graph (DAG) of agent tasks. Uses few-shot prompting with Claude Sonnet on Bedrock.
Discovers the best agent for each task using semantic search over capability embeddings stored in PostgreSQL with pgvector.
Executes the plan by dispatching tasks to agents via Redis queues, tracking progress, and aggregating results in real time.
Each Agentic Unit is a self-contained microservice with its own model, manifest, and task queue
GPU-accelerated document OCR using DeepSeek vision models. Extracts text from scanned PDFs and images.
Multi-format document parsing for PDF, DOCX, and plain text. Structures content for downstream agents.
LLM-based evaluation and fraud detection. Scores document quality and flags anomalies.
Generates structured Markdown reports from processed data. Summarizes findings across agent outputs.
Text-to-speech audio report generation. Creates audio summaries using Amazon Polly on AWS.
Visual summary generation using Bedrock image models. Creates infographic-style representations.
PostgreSQL with text-to-SQL. Stores structured data and enables natural language queries over results.
EU VAT number validation via the VIES API. Verifies business registration and tax compliance.
Web search and content extraction. Enriches workflows with external data from the web.
Design decisions that make AOA production-ready
Each AU encapsulates a focused capability with its own model, manifest, and task queue. Agent Cards capture NFR metadata — performance, reliability, cost — enabling fitness-based discovery, not just keyword matching.
Not every task needs a frontier model. AOA uses the smallest model that gets the job done — cloud LLMs for planning, specialized models for OCR, simple APIs for validation. Cost-efficient without sacrificing quality.
Three layers of memory emerge from the architecture: planner DAGs as episodic memory, registry embeddings as semantic memory, and orchestrator logs as working memory. The system learns from what it has solved before.
Agent logic is decoupled from model inference. The same agent code runs on a DGX Spark with local GPU models or on AWS with Bedrock APIs — just change an environment variable. No code changes needed.
Production-grade infrastructure, all containerized